A Dual-Path Transformer with Discriminative Region Mining for Fine-Grained Plant Pest Recognition

Accurate identification of plant pests is essential for maintaining agricultural productivity and ensuring food security. However, the high visual similarity between pest species, small target size and cluttered backgrounds in field images pose significant challenges for fine-grained pest recognition. In this study, we propose a dual-path Transformer model, termed Global-Local Aligned Mining Vision Transformer (GLAM-ViT), designed to enhance recognition performance through discriminative region mining and multi-scale feature integration. The model incorporates a semantic-guided patch selection mechanism that computes the similarity between the class token and patch tokens to locate high-response regions, which are then used to generate fine-grained views through adaptive cropping. These views are encoded in a separate path sharing the backbone with the original image, enabling complementary representation learning. Furthermore, a lightweight feature aggregator fuses global and local semantic cues, and a class-center constraint loss based on cosine similarity is introduced to optimize inter-class separability and intra-class compactness. Extensive experiments conducted on two benchmark datasets, IP102 and WPIT9K, demonstrate that the proposed method consistently outperforms existing CNN and Transformer-based models in accuracy and robustness. These results confirm the effectiveness of the dual-path structure and discriminative region guidance in addressing the unique challenges of fine-grained pest recognition in complex agricultural environments.